Abstract
This paper compares the performances of three supervised machine learning algorithms in terms of predictive ability and model interpretation on structured or tabular data. The algorithms considered were scikit-learn implementations of extreme gradient boosting machines (XGB) and random forests (RFs), and feedforward neural networks (FFNNs) from TensorFlow. The paper is organized in a findings-based manner, with each section providing general conclusions supported by empirical results from simulation studies that cover a wide range of model complexity and correlation structures among predictors. We considered both continuous and binary responses of different sample sizes. Overall, XGB and FFNNs were competitive, with FFNNs showing better performance in smooth models and tree-based boosting algorithms performing better in non-smooth models. This conclusion held generally for predictive performance, identification of important variables, and determining correct input-output relationships as measured by partial dependence plots (PDPs). FFNNs generally had less over-fitting, as measured by the difference in performance between training and testing datasets. However, the difference with XGB was often small. RFs did not perform well in general, confirming the findings in the literature. All models exhibited different degrees of bias seen in PDPs, but the bias was especially problematic for RFs. The extent of the biases varied with correlation among predictors, response type, and data set sample size. In general, tree-based models tended to over-regularize the fitted model in the tails of predictor distributions. Finally, as to be expected, performances were better for continuous responses compared to binary data and with larger samples.
Introduction
In high-stakes industries like banking and finance, tabular (structured) data remains the standard. While practitioners have increasingly adopted complex "black-box" machine learning algorithms over traditional linear models, a critical gap remains: trust. In regulated environments, understanding why a model made a specific prediction is often as important as the prediction itself. This paper investigates whether popular interpretability tools—such as Partial Dependence Plots (PDPs) and Variable Importance (VI)—accurately reflect the underlying "ground truth" or if they potentially mislead researchers.
Methodology
The authors conducted a large-scale simulation study, allowing for a direct comparison of model outputs against an "Oracle" (the known true functional form). The study evaluated three primary algorithms: scikit-learn implementations of XGBoost (XGB) and Random Forest (RF), and TensorFlow Feedforward Neural Networks (FFNN). To ensure each model reached its peak performance, the researchers employed rigorous hyperparameter optimization (HPO) strategies.
Data Generation and Functional Forms
The core of the empirical analysis relied on 12 distinct mathematical structures, ranging from simple linear models to complex "jumpy" surfaces with local interactions. These functional forms represent the ground truth that the models attempted to recover.
Datasets were generated with varying sample sizes (50,000 and 500,000 observations), predictor sets (up to $p=50$), and correlation levels ($r = 0, 0.25, 0.5$) to observe how feature dependency impacts model interpretation.
Key Insights and Results
Smooth vs. Jumpy Relationships
There is no universal winner among algorithms. FFNNs, utilizing ReLU activation, effectively act as local linear models, making them superior for smooth, continuous relationships. Conversely, XGBoost’s piecewise constant nature makes it the preferred choice for "jumpy" data characterized by sharp turns or discontinuities.
The Random Forest Flaw
Random Forest was the clear underperformer in this study. It overfitted training data more significantly than its counterparts and struggled to capture complex interactions. The authors suggest that standard hyperparameter tuning often fails to make RF trees deep enough to uncover the full underlying model structure.
The "Squeezing" Effect and Interpretability Biases
A major finding was that tree-based models (XGB and RF) tend to over-regularize in the "tails" of predictor distributions. This "squeezing" effect means that PDPs often underestimate the true relationship at the edges of the data range. Furthermore, high correlation among predictors acts as an "interpretability killer," leading to "ghost interactions" in CICE and PDP plots where the model appears to find relationships that do not exist in the ground truth.
Conclusion
For practitioners, the message is clear: do not trust interpretability plots blindly. While XGB and FFNN are highly competitive in predictive power, the "truth" shown by a PDP is filtered through the specific architecture of the chosen algorithm. FFNNs generally showed less overfitting and better interaction capture in smooth models, while XGB excelled in non-smooth scenarios. Increasing sample size consistently reduced bias and variance, leading to more faithful model interpretations.
Summary updated